Highly efficient YOLOV9 model for detecting extremely small-scale objects

Abstract Over the last decade, advancements in small object detection have been notable, yet existing models struggle with extremely small objects due to (1) their reduced informational significance compared to ground-view objects and (2) challenges in detecting unevenly distributed objects, particularly in dense regions. To tackle these challenges, this paper presents an efficient YOLOv9 model incorporating two innovative modules. The first, called RepNCSPELAN4, is designed to accelerate feature extraction. It takes input from the initial convolutional layer and splits it into two separate paths: one processed through RepNCSP layers and the other through standard convolutional layers. The outputs are then merged, optimizing gradient flow and feature reuse. This dual-path approach boosts learning efficiency and inference speed while maintaining low computational complexity. The second module SPPELAN incorporates Spatial Pyramid Pooling (SPP) into ELAN for layer aggregation. It employs a convolutional layer for channel adjustment, then performs spatial pooling operations to capture multi-scale contextual information. Concatenates outputs and consolidates features via another convolutional layer, optimizing detailed feature extraction across spatial hierarchies, and Experimental validation on the VisDrone dataset demonstrates superior performance, with our model achieving higher mAP (48.7%) and precision (78.3%) compared to existing models, particularly in detecting x-small items.